MHSA: A Multi-scale Hypergraph Network for Mild Cognitive Impairment Detection via Synchronous and Attentive Fusion

Manman Yuan, Weiming Jia, Xiong Luo, Jiazhen Ye, Peican Zhu, Junlin Li

TL;DR

This paper introduces MHSA, a novel multi-scale hypergraph network for detecting Mild Cognitive Impairment (MCI). It uniquely incorporates phase synchronization (PLV) from the spectrum domain and dynamic temporal connectivity, fused via an attention mechanism, to build more comprehensive brain network models for improved MCI detection accuracy.

What Problem Does This Paper Address?

Early and accurate detection of Mild Cognitive Impairment (MCI) is crucial for timely intervention and preventing further cognitive decline. While hypergraph networks have shown promise in analyzing brain networks for MCI detection, existing methods often rely on simple vector distances between brain regions. This approach has limitations: it struggles to capture the complex, dynamic, and higher-order interactions within brain networks, especially those involving synchronized activity across different brain regions. Furthermore, effectively fusing information from different scales and domains (like temporal and spectral) remains a challenge. This paper aims to address these limitations by developing a hypergraph model that explicitly accounts for synchronized brain activity and integrates multi-scale dynamic connectivity features.

Methods

The core of the MHSA framework lies in its novel approach to constructing hypergraphs that represent brain networks.

  • Synchronization Analysis Model: The paper first utilizes the Phase-Locking Value (PLV) to quantify the phase synchronization between different Regions of Interest (ROIs) in the spectrum domain of fMRI signals. This captures how synchronized the brain activity is across regions at specific frequencies.
  • Multi-scale Feature Fusion Model: To integrate dynamic connectivity, an attention mechanism is employed to assess the importance of ROIs in the temporal domain. This attention mechanism dynamically adjusts coefficients to optimize the contribution of each ROI to phase synchronization. This temporal information is then fused with the spectral synchronization information (CSPE) using a Hadamard product to create a comprehensive temporal-spectrum fusion matrix (CTS).
  • Hypergraph Construction Model: The CTS matrix is then used to construct a hypergraph. This involves defining hyperedges that connect multiple ROIs based on their synchronized and dynamic connectivity. The hypergraph is further processed using hypergraph neural networks (HGNNs) to learn rich feature representations.
  • Classification: Finally, a multilayer perceptron (MLP) is used for disease classification based on the learned hypergraph representations.

The novelty lies in the explicit incorporation of spectral phase synchronization (PLV) and the attentive fusion of temporal and spectral dynamics, moving beyond simple distance-based similarity measures.

Datasets

The study utilizes the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. This dataset contains fMRI data from 179 subjects, categorized into:

  • 51 patients with Mild Cognitive Impairment (MCI)
  • 64 patients with early MCI (EMCI)
  • 36 patients with late MCI (LMCI)
  • 28 age-matched normal controls (NC)

The data was preprocessed using the DPARSF package, including standard steps like slice timing correction, motion correction, co-registration, segmentation, normalization, smoothing, and bandpass filtering.

Key Insights and Ideas

  • Synchronization is Key: The paper strongly suggests that phase synchronization in the spectrum domain is a critical, underutilized feature for understanding brain network dynamics in MCI. Simply looking at spatial proximity or temporal correlations isn't enough.
  • Multi-scale Fusion is Powerful: Combining information from both the temporal and spectral domains, and doing so in a multi-scale fashion, leads to a more holistic representation of brain activity.
  • Attention for Dynamic Importance: The attention mechanism effectively weights the contribution of different ROIs over time, allowing the model to focus on the most relevant dynamic connectivity patterns.
  • Hypergraphs for Higher-Order Interactions: Hypergraphs are well-suited for capturing the complex, non-pairwise relationships in brain networks, which is essential for understanding intricate neurological conditions like MCI.

Key Results

The MHSA model demonstrated superior performance across various MCI detection tasks compared to several baseline methods (SVM, GCN, GAT, GCNII, GCNH, CAGNN, SIGAT, HGNN).

  • NC vs. EMCI, MCI, LMCI: MHSA consistently outperformed baselines in accuracy, sensitivity, precision, and F1-score across these specific subtype classifications (Table II).
  • NC vs. ALL TYPES of MCI: When consolidating all MCI subtypes, MHSA achieved a perfect sensitivity score of 100% and an accuracy of 92.21%, outperforming the next best model by 2.23% in accuracy and 1.36% in F1-score (Table III).
  • Ablation Studies: Experiments removing PLV or the attention mechanism (MHSA-PLV, MHSA-attention) showed a significant performance drop, highlighting the indispensable role of both components.

Limitations and Future Work

  • Limitations: The paper acknowledges that while the model shows promise, the complexity of brain networks means there's always room for improvement. The reliance on specific preprocessing steps for fMRI data might also influence results.
  • Future Work: The authors suggest exploring the application of the MHSA model to other neurodegenerative diseases. Further research could also investigate more advanced hypergraph construction techniques or explore different fusion strategies for temporal and spectral features. The paper also hints at the potential of analyzing specific brain regions identified as highly connected in MCI networks for better prediction of disease progression.

Summary updated

Abstract

The precise detection of mild cognitive impairment (MCI) is of significant importance in preventing the deterioration of patients in a timely manner. Although hypergraphs have enhanced performance by learning and analyzing brain networks, they often only depend on vector distances between features at a single scale to infer interactions. In this paper, we deal with a more arduous challenge, hypergraph modelling with synchronization between brain regions, and design a novel framework, i.e., A Multi-scale Hypergraph Network for MCI Detection via Synchronous and Attentive Fusion (MHSA), to tackle this challenge. Specifically, our approach employs the Phase-Locking Value (PLV) to calculate the phase synchronization relationship in the spectrum domain of regions of interest (ROIs) and designs a multi-scale feature fusion mechanism to integrate dynamic connectivity features of functional magnetic resonance imaging (fMRI) from both the temporal and spectrum domains. To evaluate and op-timize the direct contribution of each ROI to phase synchronization in the temporal domain, we structure the PLV coefficients dynamically adjust strategy, and the dynamic hypergraph is modelled based on a comprehensive temporal-spectrum fusion matrix. Experiments on the real-world dataset indicate the effectiveness of our strategy. The code is available at https://github.com/Jia-Weiming/MHSA.
BibTeX
@inproceedings{yuan2024mhsa,
  author = {Manman Yuan and Weiming Jia and Xiong Luo and Jiazhen Ye and Peican Zhu and Junlin Li},
  title = {{MHSA: A Multi-scale Hypergraph Network for Mild Cognitive Impairment Detection via Synchronous and Attentive Fusion}},
  booktitle = {BIBM},
  year = {2024},
  doi = {10.1109/bibm62325.2024.10822410},
}
Figure